1

John Zeimbekis

Seeing, Visualizing, and Believing: The Cognitive Penetrability of Picture Perception

Abstract. Visually ambiguous figures have always been used to test claims about the cognitive

impenetrability (CP) of certain visual processes and about how to draw the perception-cognition

distinction (Fodor 1988; Churchland 1988; Pylyshyn 1999, 2003; Macpherson 2006;

Raftopoulos 2009, 2011). I distinguish several kinds of visual ambiguity found in picture

perception (§1), and focus on two which have not been discussed previously in the CP literature

but pose a significant threat to claims that vision is cognitively impenetrable. One is the ability to

visually experience a picture surface as flat after it has caused volumetric nonconceptual contents

(§§2-3); the other is the ability to use a surface initially perceived as flat to visualize three-

dimensional scenes (§4). I argue that these forms of visual ambiguity emerge because in picture

perceptions, the visual processes that construct egocentric volumetric representations (equivalent

to Marr’s 2½D sketch) have to rely solely on monocular depth cues in the absence of parallax

and stereopsis. This gives perceivers a degree of personal-level control over which visual

experiences they have (2½D or 2D). However, in the final section (§5), I hold that visual

contents resulting from such penetration do not cause perceptual beliefs, and offer an account of

pictorial illusions (trompe-l’oeils) as rule-proving exceptions to this claim. The result is that the

form of penetration involved does not have the pernicious epistemological consequences usually

expected of CP.

2

I. Pictures and visual ambiguity

Among theories of depiction, there is considerable consensus that pictures can cause not only

visual experiences with volumetric contents but also veridical visual experiences of picture

surfaces as flat. The assumption is so widespread that it splits into several varieties, depending

on whether viewers can have both visual experiences at once. One variety holds that it is possible

to have both the experience of volumetric shapes and the experience of flatness at once

(Peacocke 1987, 386, 394; Schier 1986, 9; Lopes 1996, 40; Walton 1990, 293-304). Another

says that we always have both experiences at once (Wollheim 1987, 46; Gibson 1986, 282; Rock

2001, 98). A third holds that we cannot have both experiences at once, but can have each one

separately (Gombrich 1960, 224, and perhaps Cutting and Massironi, 1998). But there is one

point on which all of these authors agree: that we are capable of having both kinds of visual

experience, both of flatness and of depth, of picture surfaces. If this consensus on the

phenomenology of picture perception is anything to go by, pictures qualify as visually

ambiguous figures. The consensus is worth noting because, as Macpherson has said in a related

context,1 ‘the evidence that one must appeal to here is introspective and the more people that

agree that a change takes place in their experience the better’ (Macpherson 2006, 91).

1 The context is Macpherson’s discussion of the ability to perform Gestalt switches with

ambiguous figures; but in the cases Macpherson has in mind the switch is not from 3D (2.5D, ie,

egocentically 3D) to 2D visual contents, as will be the case with the examples introduced here.

3

Yet, if we admit the phenomenological claims as a form of evidence, then pictures should

support not one but several kinds of visual ambiguity. In one kind, visually experiencing a

picture (or even parts of it at a time) as a flat surface (one perpendicular to the line of vision)

requires great mental effort. Figure 1 provides an illustration. If we can visually experience such

pictures or parts of them at a time as flat, it will only be subsequent to the pictures’ having

caused visual experiences with volumetric content; the mental effort could involve consciously

shifting attention across the surface until we find a way to blind ourselves to regions that work as

depth cues and prevent the cues from triggering depth representations. Even Peacocke’s

relatively modest claim, that we experience the outlines of object-representations in pictures as

flat,2 seems wrong where this picture is concerned: the left outline edges of the pyramids seem

further away than the right edges. The problem is that vision refuses to process just the outlines;

it processes the interior lines as edges and yields a mental representation of a regular volume

with a certain orientation, which settles the egocentric distances of the object’s parts, including

parts of the outlines.

2 “The silhouette both is and is experienced as a flat surface, or at least as occupying a plane: any

description of the experience omitting this point is incomplete.” Peacocke 1987: 386.

4

Figure 1

Whether or not figure 1 is in fact visually ambiguous (as one would expect given the claims that

pictures can also cause visual experiences of their surfaces as flat), it is clear that when we

perceive it we first become aware of volumetric visual contents and then have trouble

suppressing this visual experience by making an agentive mental effort. How do the volumetric

visual contents of such ‘natural’ picture perceptions emerge in the first place? We could sketch

the following hypothesis, postponing until the next section its details and justification. The

picture’s surface includes monocular depth cues which can trigger brain representations of

volume even without calculations from parallax and stereopsis; those cues exploit early visual

processes dedicated to constructing volumetric contents; the outputs of those early processes are

brain representations corresponding to Marr’s 2.5 sketch; those representations provide the

contents of phenomenal awareness; and those phenomenal contents constitute the visual

experience of the picture’s content. If the hypothesis is plausible, as I will try to show in section

2, then, in this form of visual ambiguity, volumetric contents are experienced without any help

from visualizing or visually imagining, which can be driven consciously and are usually thought

5

to be higher-level processes. (See Tye 1991, 90, 96; Pylyshyn 2007, 141; Pylyshyn 2003;

Kosslyn 1994.) On the other hand, 2D contents emerge – if they emerge at all, or when they do3

– only subsequently, as a result of some conscious activity.

Today we frequently have such picture perceptions, which feel effortless and automatic, because

many of the pictures we perceive (including non-agentive pictures like photographs) contain cues

which successfully exploit object-perception processes like segmentation and the construction of

egocentric depth representations out of two-dimensional distributions of light. But while some

depth cues (line junctions) were mastered in prehistoric times (Cavanagh 2005, Biederman and

Kim 2008), it took a long time for drawing and painting to develop techniques which could cause

visual and perceptual contents comparable to those caused by objects about shading, textures,

colours, lines, orientations, volumes, and the depth relations between objects (see, for example,

Kubovy 1986 on depth relations; Gombrich 1960, Ch. 1, on colour). Moreover, once mastered,

the techniques were often subsequently avoided by painters trying to develop new pictorial styles

– that is, new ways of causing a visual impression of depth and volume without relying on

existing techniques. When a picture does not include adequate cues to evoke some of those

features (shading, textures, lines, orientations, etc), the brain cannot form representations of the

intended picture contents solely on the strength of stimulus-driven or hard-wired processes. On

the strength of those processes, the brain may represent the picture as a flat surface with uneven

stains, or else it may capture only some local depth features (like apparent occlusions) but not

global ones or depth relations between objects. In those cases, we have to perform conscious acts

of visualizing by using the picture surface as a prop.

3 I think they do emerge in some cases; see section 3.

6

This gives rise to a second kind of visual ambiguity supported by pictures. This time, the surface

naturally yields a visual experience of a flat surface; but subsequently, by performing conscious

acts of visualizing, we can look at the surface and generate experiences of depth and volume.

Such cases of depiction are often provided by what Cutting and Massironi (1998) call fortuitous

pictures: natural or other objects whose colour and line patterns can, but do not always, support

visualizing activities which allow ‘seeing-in’. The passage below, taken from Leonardo’s advice

to apprentice painters, seems to describe this kind of ambiguity:

if you look at any walls soiled with a variety of stains, or stones with variegated

patterns, when you have to invent some location, you will therein be able to see a

resemblance to various landscapes graced with mountains, rivers, rocks, trees,

plains, great valleys and hills in any combinations. Or again you will be able to see

various battles and figures darting about, strange-looking faces and costumes, and

an endless number of things which you can distill into finely-rendered forms. And

what happens with regard to such walls and variegated stones is just as with the

sound of bells, in whose peal you will find any name or word you care to imagine.

(Leonardo 1989, 222)

For instance, it is possible to visually experience figure 2 as the stained surface that it is:

7

Figure 2

But it is also possible to visually experience the figure depthwise as figure and ground; and in

fact concepts can help us to settle the figure-ground relation, for instance if we’re told to see the

lower part as a field of wheat and the upper part as a grey sky. A question that emerges about

such cognitively driven acts of visualizing is whether they just add a cognitive ingredient to

mental contents without altering the outputs of the visual processes leading up to Marr’s 2.5D

sketch, or instead tamper with the way vision itself generates figure-ground segregations; I

postpone this discussion until section 4.4

A third form of picture-related visual ambiguity occurs when neither 2D nor 2.5D visual contents

are the more natural output of vision, so that neither prevails in perception. Details of pictures

taken out of context can be ambiguous in this way because their depth cues can easily be

overridden in isolation. Many Rorschach blots are also in this category: they cause

representations of volume only indirectly (through recognition, by exploiting templates for

outlines); they do not include shading and enclosed lines that could signal convexities,

concavities or what Marr calls surface orientation discontinuities (Marr 2010, 215-233). The

weakness of Rorschach blots as pictures is described by Gibson when he writes that their

4 The figure is a detail of Turner’s Approach to Venice, 1844; Andrew W. Mellon Collection.

8

‘invariants are all mixed up together and are mutually discrepant instead of being mutually

consistent or redundant’ (Gibson 1986, 282). A fourth kind of ambiguity is bi-stability in which

both visual interpretations are volumetric and both are generated bottom-up, as apparently occurs

when we switch among different perceptions of the Necker cube. It seems plausible to hold, as

Pylyshyn does (2003, Ch. 1), that both visual experiences of the Necker cube are performed

purely by bottom-up processes jointly with hard-wired processes. Finally, a fifth kind of

ambiguity is that found in perceptions of the Jastrow drawing. Both experiences caused by

Jastrow’s drawing are experiences of the picture’s content, namely, an object belonging to a

three-dimensional kind. However, if those experiences involve any representation of volume,

that representation depends almost entirely on prior knowledge encoded in sortal concepts, not

stimulus-driven depth cues. In fact, it is not immediately clear that there is any change of visual

contents in the Jastrow’s ambiguity, although this has been claimed recently (see Macpherson

2006, 97; Siegel 2011 makes a related point, discussed here in section 5). In the Jastrow figure,

representations of volume and depth are not caused by bottom-up or hard-wired visual processes,

as they are by the Necker cube, because the Jastrow contains no internal lines suggestive of

surface discontinuities – only a couple of weak volumetric cues from shading (the shading that

suggests a concavity between the head and neck). So any change of visual experience as we

switch from one experience to the other is unlikely to be a change in content construed as what

would have to be the case in the world for the experience to be true. The change seems to be

more superficial; it has been described by Lyons as ‘manipulating the representation produced’

previously by the bottom-up processes without altering it, or as ‘facilitating pop-out of certain

patterns’ and yielding ‘a late experiential effect, leaving the nonconceptual early perceptual

states unaffected but influencing the nondoxastic seemings’ (Lyons 2011).

9

While cases belonging to all five kinds of ambiguity have been described and discussed in the

literature on depiction, the kinds of visual ambiguity supported by pictures have not yet been

distinguished, nor the differences between them explained. Making the distinctions is important

for two sorts of reasons. First, ambiguous figures have always been used to test claims about the

cognitive impenetrability of certain visual processes and certain ways of drawing the perception-

cognition distinction (Churchland 1988, Macpherson 2006). Deniers of cognitive penetrability

(Fodor 1988, Pylyshyn 1999, Raftopoulos 2009, 2011) have in turn responded to each kind of

ambiguity-based penetrability claim. Some kinds of ambiguity are easy enough for modularists

to deal with, others less so. But pictures are an extremely rich resource of different kinds of

visual ambiguity which has barely been used. I have isolated the two kinds which I think pose

the greatest threat to impenetrability claims: the ambiguity of natural picture perceptions and that

of cognitively driven picture perceptions. Secondly, understanding the different forms of visual

ambiguity is essential to understanding what pictures are. For example, they show that picture

perception is not a single kind of mental state, be it a higher-level state of visualizing or a lower-

level state which is the output of earlier visual processes;5 they allow us to describe the mental

states and contents caused by different kinds of pictures, and thus to relate those states and

contents to other states such as perceptual belief and memory; they allow us to account for what

is called naturalism in depiction; and so on. Thus, issues about the perception-cognition

5 For example, Abell and Currie (1999, 440), Wollheim (1987, 1998) and Levinson (1998, 232)

seem to be committed to an account of picture perception as the output of pre-doxastic processes;

Walton (1990) to a doxastic-level account (though see Walton 2002, 31, for a denial of this).

Most authors on depiction do not address the issue at all; an exception is Levinson 1998.

10

distinction turn out to be essential for an account of depiction; though in this paper I will focus

on the first set of issues.

The first two kinds of visual ambiguity described correspond to two kinds of picture perception.

The first, illustrated by figure 1, could be called bottom-up or natural picture perception. The

ambiguity it gives rise to is one in which an object naturally causes 2.5D visual contents, but

with some mental effort we can also see the object as flat. The second, illustrated by figure 2,

could be called cognitively driven picture perception; the ambiguity it gives rise to is one in

which an object naturally causes 2D visual contents, but with some mental effort we can see the

object as 2.5D.

How exactly do these two kinds of visual ambiguity relate to the cognition-perception distinction

and to cognitive penetrability? If all picture perceptions were bottom-up and hard-wired, or if all

picture perceptions were cognitively driven acts of imagining, or if some were bottom-up and

others cognitively driven, this would not necessarily entail any form of cognitive penetrability.

The bottom-up picture perceptions could be compatible with impenetrability of the visual

processes up to Marr’s 2.5D sketch, while the cognitively driven ones could conceivably be such

that they added a cognitive phenomenology to mental contents without altering the outputs of the

visual processes leading up to Marr’s 2.5D sketch. It is along similar lines that Raftopoulos

(2009, 2011), Fodor (1988) and Pylyshyn (2003) have dealt with the kinds of ambiguity

supported by the Jastrow figure and the Necker cube. But there is a difference between those

forms of visual ambiguity and the forms described above which involve switching from 2D to

2.5D visual contents and vice versa: only the latter two have the potential to cause trouble for the

11

hypothesis that visual processes required to construct 2.5D nonconceptual content out of the 2D

sketch are cognitively impenetrable.

My strategy in this paper is twofold. I will concede that there are good counterexamples to the

claim that visual processes leading up to Marr’s 2.5D sketch are cognitively impenetrable.

Sections 3 and 4 describe two forms of visual ambiguity and argue that they imply the

cognitively penetrability of visual processes required to construct 2.5D content out of the 2D

sketch. I believe that trying to exclude all such cases of penetrability would be barking up the

wrong tree. Instead – and this is the second part of the strategy – I will claim that even if it turns

out that the functioning of some key visual processes can sometimes be tampered with by

processes that count as cognitive, this does not lead to the expected pernicious epistemic and

epistemological consequences (sections 5 and 6). The reason is that visual experiences caused by

pictures fail to cause perceptual beliefs. This also amounts to something of an enabling condition

for depiction as a form of representation, as opposed to a source of illusions.

2. Bottom-up or natural picture perceptions

The hypothesis that some pictures cause visual experiences with 2.5D contents without any

conscious mental contribution on behalf of the viewer is plausible, because otherwise picture

perception would always be an effortful, slow performance for subjects, involving acts of

visualizing for each feature of the picture that can be used to imagine a depth relation or a

volumetric shape. Instead, most picture perception is caused in an effortless, quick and automatic

way, suggesting that it is subserved by hard-wired and bottom-up processes, not deliberate acts

12

of imagining. So the hypothesis is compatible with the phenomenology of many picture

perceptions; and it also allows us to explain the other cases: in the absence of cues that trigger

automatic, dedicated processes, we have to perform slow, attention-consuming acts of

visualizing to grasp the picture contents. A sketch follows of how bottom-up picture perceptions

could occur. It focuses mainly on the construction of volumetric shapes from monocular cues,

and for reasons of space neglects the segregation of objects during picture perception and the

ways in which different object-recognition processes are co-opted by different kinds of pictures;

issues which are discussed in detail in my 2012a.

The natural place to look for support for the claim that many pictures cause egocentric, conscious

representations of 3D shapes is Marr’s theory of vision. Marr’s working hypothesis is that initial,

two-dimensional retinal inputs are built up into adequate representations of the three-dimensional

objects that cause them, and this makes its details particularly interesting for a theory of

depiction. An important part of Marr’s proposal cannot be applied to pictures. It is the part of his

theory (Marr 2010, Ch. 3) according to which binocular disparity, the slight difference in

perspective from which each eye views the object, is exploited by the visual system to compute

the relative depths of surfaces in a scene. In pictures, all the points of the surface from which the

eyes receive stimuli are at roughly the same location on the back-to-front axis (the axis

perpendicular to the picture’s surface and passing through the centre of the viewer’s body, which

is identical to the z axis of Peacocke’s nonconceptual scenario content; see Peacocke 1992, 62).

Therefore, in picture perception, the brain cannot compute depth relations from binocular

disparity and has to rely on monocular depth cues. The same applies to parallax from movement

relative to the object. When we move relative to the picture, points on the surface are not seen as

13

moving at different speeds and parallax cannot be exploited to compute depth. Picture perception

is even insensitive to changes in the angle from which the picture is viewed; the picture can be

rotated up to 22 degrees on its vertical axis without this deforming the three-dimensional

representations it causes (Cutting 1987), an effect Kubovy (1986) calls ‘robustness of

perspective’.

However, Marr also relies heavily on hypotheses about how we extract volumetric and depth

information from monocular cues which exploit neither binocular disparity nor parallax. Here,

his proposal is that parts of the visual system take two-dimensional representations of objects or

scenes as inputs, apply hard-wired processes to them, and yield three-dimensional shape

representations. The inputs of the processes are edges (lines) and their orientations and junctions,

that is, the outputs of earlier visual processes which generate a brain representation of differences

in light intensities. These are treated by vision as discontituities in the distance of surfaces from

the viewer (‘occluding contours’), convexities or concavities of surfaces (‘surface orientation

discontinuities’) and depthwise curvature (‘surface contours’) (2010, 215-233). Texture patterns

are also interpreted to yield representations of the depthwise slant of a surface (233-239).

According to Marr, the brain applies a set of simplifying constraints when interpreting occluding

contours; for example, points close on the 2D contour are assumed to be close in 3D space (223),

an assumption which is nicely illustrated by Richard Gregory’s impossible triangle. When those

assumptions are applied to two-dimensional views of certain regular 3D shapes,6 they yield

6 The regular shapes are ‘generalized cones’, shapes whose cross sections have the same shape

but can vary in size (for example, a sphere, a pyramid or a cylinder). Although the objects we

have to recognize in the visual environment are not usually generalized cones, they can be

14

accurate three-dimensional representations of those shapes. (Note that monocular cues are likely

to be an essential resource for vision when objects are too distant for binocularity or parallax to

calculate the relative distances of parts of an object.)

On Marr’s model, the output of these processes – the 2.5D sketch, an egocentric volumetric

representation – subserves kind-recognition and therefore precedes the application of concepts.

Volume is assigned as a condition for recognition and classification. The opposite –

classification as a condition for assigning shape – would amount to a form of sortalism like the

one criticized by Campbell (see his account of the ‘delineation thesis’; Campbell 2002, 69). On

the sortalist scenario, we would carve up the visual scene into volumetric objects by using sortal

concepts, not on the basis of bottom-up stimuli jointly with hard-wired visual processes like

those that Marr describes; those processes would under-determine the assignment of shape,

allowing different shapes to be assigned depending on which concepts we applied.

A sceptic might argue that in picture perception, due to the absence of any differential stimuli

from stereopsis and without performing the processes that compute them, the brain would not be

in the kinds of states that qualify as causal antecedents for having three-dimensional object

representations. While we do have 3D object representations at some point in picture perception,

the sceptic might argue that they are not caused by lower-level visual processes but are instead

the effects of some kind of higher-order imagining or visualizing. My response to this scepticism

is twofold. First, I do not deny that there are cognitively driven picture perceptions which

analysed coarsely into regular component shapes for the purposes of kind-recognition (Marr

2010, Biederman 1987, 1995; see also the brief discussion below).

15

involve conscious acts of imagining. Such picture perceptions are different to the ones I call

‘bottom-up’ or ‘natural’, and the sceptic would have to explain the differences. Admitting

bottom-up picture perceptions can explain the difference between perceiving figure 1 and

perceiving figure 3:7

Figure 3

According to its title, Figure 3 represents a guitar player, and it does allow us to use it to visually

represent a guitar player – at least if the concept of visual representation is construed to include

the contents of visualizing. But this requires conscious hypothesizing about depth relations and

volumetric shape. Before we read the picture’s title and make use of concepts, all we can only

pick up in terms of volume or depth are, at best, some local occlusion effects which do not allow

us to reconstruct the volumes corresponding to the object described by the sortals. Figure 1 on

the other hand yields that kind of information naturally and effortlessly as soon as we glance at

the picture, suggesting that Marr is right to hold that the brain can reach volumetric object

7 Picasso, The Guitar Player, 1910. Musee National d'Art Moderne, Paris.

16

representations without either the benefit of binocular disparity or input from recognitional

concepts.

Object-recognition theories after Marr seem to confirm this hypothesis. According to one such

theory, Biederman’s theory of recognition by components (‘RBC’; Biederman 1987, 1995), the

features of objects on the basis of which we analyse them into components, such as vertices,

which are interpreted as concavities or convexities, are ‘generally invariant over viewing

position’ (1987, 115). As a result, the information required to analyse an object into primitive

volumetric components can be extracted from a single two-dimensional representation of the

object (1995, 153; 1987, 133-141). A picture of an object causes a two-dimensional view of that

object from a single viewpoint, so on this theory, picture perceptions can suffice to cause

volumetric object representations. Note that on RBC theory, these structural representations are

object centered, not viewpoint relative; they correspond to the objects in Marr’s 3D sketch of the

visual scene and are thought to be the immediate causal antecedents of object recognition on this

theory as well as Marr’s.

Competing views of object recognition have shown that RBC is inadequate for explaining certain

recognition tasks and fails to match a quantity of experimental data. The competing accounts are

called ‘image-based’, ‘viewpoint-relative’ or ‘multiple-views’ theories. Image-based or

viewpoint-relative models of object recognition emerged from evidence that recognition also

relies on viewpoint-relative information in ways that RBC does not account for (Tarr and Pinker

1989, Bulthoff and Edelman 1992, Edelman and Bulthoff 1992, Tarr 1995, Ullman 1998). In a

series of experiments, subjects who were acquainted with new kinds of objects could still only

17

recognize them from viewpoints similar to those under which they were presented to them,

contradicting the predictions of RBC theory. (This behavioural evidence has also been backed by

neurological findings; see Logothetis et al. 1995.) Moreover, RBC’s structural representations

are too coarse-grained to capture the recognition of individuals, or even to subserve many fine-

grained generic classifications. These theories could be seen to present a challenge to the idea

that a volumetric (egocentric, 2.5D) representation is constructed prior to recognition, since they

could be compatible with the claim that volumetric representation is added cognitively post-

recognition, not visually prior to recognition. However, viewpoint-dependent theories do not

make the claim that the brain representations that trigger recognition are representations of

objects in two dimensions. The relevant distinction is that between structural, object-centered

representations, and representations that are relativized to viewpoints. The representations in

Marr’s 2.5D sketch are also relativized to viewpoints, yet they are volumetric and represent

depth; they contrast with subsequent brain representations which permit mental rotation, belong

to the 3D sketch, and qualify as structural or viewpoint-independent. The conclusion that both

viewpoint-relative and viewpoint-independent mechanisms are involved in recognition is

supported by neurological findings on an area of the human brain called the lateral occipital

complex (LOC). According to Kourtzi et al. (2003), this area represents perceived object-shape,

a relatively high-level cognitive representation, and plays a role in object recognition. Kourtzi et

al. show that while one subregion of the LOC represents two-dimensional shape, another

subregion encodes the represented three-dimensional shape of objects and appears to ‘mediate

object and scene recognition based on rather abstract three-dimensional representations’ (Kourtzi

et al., 2003, 918; see 911 for other studies which show that the LOC represents shape three-

dimensionally).

18

Volumetric representations on Marr’s hypothesis and on the RBC hypothesis, and viewpoint-

relative representations on the Tarr’s, Bulthoff’s and Ulmann’s hypotheses, are considered causal

antecedents for object recognition and classification, and therefore must be nonconceptual states.

Putting aside for a moment the non-naturalistic and fortuitous pictures that require agentively

driven attention and conceptual input to contribute to depth interpretation, and focusing for on

the important category of naturalistic pictures, this yields the following result. Descriptions of

the nonconceptual contents of perception by Evans (1982), Peacocke (1992), and recently

Raftopoulos (2009), are consistent with the descriptions Marr gives of the 2.5D sketch. Both are

perception-dependent, unlike conceptually encoded representations; both are spatially

egocentric; and in both cases, shape, orientation, texture and colour (including shading)

information is more fine-grained than prior conceptually encoded information about such

properties. (For examples, see Marr’s illustrations of how we visually experience flat surfaces

that contain monocular depth cues; 1982, 215-239.) On several current theories of the neural

correlates of consciousness, the kinds of shape representations which according to Marr

constitute the 2.5D sketch require local recurrent processing – a kind of brain state that occurs at

approximately 100 to 120 milliseconds after stimulus onset and precedes personal-level

awareness – which is thought to be the neural correlate of phenomenal awareness in visual

perception (Block 1990, 1997, 2005; Lamme 2000, 2003, 2005; Raftopoulos 2009, 32-39). If

that hypothesis is true, then the egocentric volumetric shape representations of Marr’s 2.5D

sketch constitute the nonconceptual contents of visual perception.

19

Thus, the sketch of bottom-up picture-perceptions that I have drawn here would explain an

important part of the phenomenology of picture perception: by the time we reach personal-level

awareness and the ability to think about the picture, the picture has already caused a visual

experience of a three-dimensional scene. This conclusion seems accurate phenomenologically;

for a great number of pictures, picture perception is not the result of any mental effort but occurs

naturally. However, the bottom-up account of many picture perceptions also raises new

questions: if there are bottom-up picture perceptions, then what differentiates those experiences

from perceptual illusions, and especially from the visual experiences caused by trompe l’oeils?

These questions are dealt with in section 5.

3. Picture perception and cognitive penetrability

As stated in section 1, there is considerable consensus among depiction theorists that picture

perceptions can support both 2.5D and 2D visual contents. An easy way to account for that

insistence would be to hold that such claims of visual ambiguity are limited to cognitvely driven

picture perceptions and do not apply to bottom-up picture perceptions. This response would

sidestep the underlying issue of how viewers could avoid having contents caused by bottom-up

processes. The problem with the response is that the theories tend to claim that pictures generally

– not just pictures that require cognitively driven forms of picture perception – can support such

visual ambiguity, and rely on naturalistic pictures such as Constable’s paintings. Consider figure

4, which is similar to one of the drawings Marr uses to illustrate the interpretation of surface

orientations as volumetric shapes by early, hard-wired visual processes. The natural visual

20

interpretation of the figure yields a nonconceptual representation of a cube seen from above, with

point A in front of point B.

Figure 4

Marr’s explanation of the experience of depth is this: ‘If the occluding contour shown with thick

lines is present on its own, one perceives a hexagon. The interior lines change it into a cube,

since they suggest that the occluding contour is not planar’ (Marr 1982/2010, 221). But although

it is initially difficult to perceive this figure and visually represent it as flat, it is possible. It

suffices to be told to look for three rhomboids, or for a regular hexagon, or to see the enclosed

lines as radii. If you see the hexagon, attention is distracted from point A to the perimeter and

you experience a short-lived change of visual experience. Note that figure 4 is not bistable: we

return naturally to the cube-perception once we relax the effort of attention, so there is one visual

interpretation we could call natural. Natural picture perceptions share this characteristic with

what Kanizsa and Gerbino call ‘amodal completions’, as distinct from what they call

‘represented’ or ‘merely thought’ completions, which work like cognitively driven picture

21

perceptions. (For examples of such revisable amodal completions see Kanizsa and Gerbino,

1982, Figures 9.3.c and 9.6.a.)

There appear to be two different ways to go about seeing the drawing as flat. The first is to

consciously direct attention away from the drawing’s enclosed lines and on to the perimeter,

optionally by using the concept rhomboid or hexagon to focus attention differently. That way, we

can avoid performing the hard-wired interpretation of the intersection of the three enclosed lines

as three-dimensional. Conscious manipulations of attention are a personal-level activity even

when they are not concept-driven; when they are concept-driven (eg because we are given

hexagon or rhomboid), the concept determines where attention is directed.

But there is a second way to get the same effect, which this time does not require us to

attentionally ignore the enclosed lines in Marr’s cube: we can just see them differently, as

radiating from the centre of the hexagon. In that case, we focus on the very part of the picture

that contains its depth cue and still get a 2D representation. If that is so, and if we appeal to the

attention-shift argument to explain the change in visual experience and content, we’re exposed to

the rejoinder that viewers directly interfere with the processes that construct volumetric

representations. The rejoinder may or may not be correct, but as it stands it pits one introspective

claim against another. If the dispute between deniers and defenders of penetrability hinges on

such a conflict, then it seems to reach a stalemate. This is important because the key argument by

deniers of penetrability where ambiguous figures are concerned is that shifts in where spatial

attention is focused on the scene change which data is processed, without changing the way such

data is processed (Fodor 1988; Pylyshyn 1999; Raftopoulos 2009, 2011). If Marr’s opaque cube

22

is visually ambiguous in the way described, then it is a clear counterexample to the attention-

shift argument for a key part of perception: the construction of 2.5D visual contents from lines of

the optic array that signal surface orientation discontinuities and convexities.

Which other strategies are available to block the conclusion that the switch from 2.5D to 2D

visual contents implies the penetrability of visual processes by agentively driven processes? For

one kind of ambiguity seen in section 1, that found in Jastrow’s rabbit-duck figure, it can be

argued that what changes each time is only the conceptual representation caused by the

nonconceptual visual contents, and that the visual experience and contents remain the same

throughout. That argument has been used to counter Churchland’s use of perceptual ambiguity to

challenge Fodor’s (1983) claims about the impenetrability of visual processing modules.

Churchland writes that we can make figures like the Jastrow ‘flip back and forth at will between

the two or more alternatives, by changing one’s assumptions about the nature of the object or

about the conditions of viewing,’ concluding that ‘some aspects of visual processing, evidently,

are quite easily controlled by the higher cognitive centers’ (Churchland 1988, 172). Discussion

of the Jastrow figure received a new twist when Macpherson (2006) claimed that in Jastrow-type

ambiguity, the nonconceptual content of visual experience also changes. (A distinct but

substantially similar position is Siegel’s 2010 generalized claim that conceptual content

influences perceptual content.) Irrespective of whether those claims are true and what kind of

penetrability they amount to, the claim that in figure 4 the switch is merely conceptual cannot

even get off the ground. It is the nonconceptual representation of shape, in particular the 2.5D

sketch itself, that changes. For example, points A and B go from being at different distances on

23

the z axis of Peacocke’s nonconceptual scenario content to being equidistant on that axis. There

is no comparable modification of nonconceptual content in the the Jastrow figure’s ambiguity.

A related strategy for countering cognitive penetrability claims is to use low-level perceptual

illusions like the Muller-Lyer illusion (Fodor 1988; Pylyshyn 1999, 2003) to argue that the

modules that yield visual experiences as outputs cannot be influenced by belief. But the Muller-

Lyer illusion, like the chessboard illusion, are not visually revisable. In the chessboard illusion,

the visual processes that interpret changes in light intensity in order to preserve light constancy

turn out to be impenetrable. We cannot as it were ‘reach down’ from personal-level awareness

into the brain processes that interpret light intensity data, tamper with them, and produce an

experience of a different colour. However, we do seem to be able to do something like this with

certain processes that yield volumetric representations, since for example when we switch from

2.5D to 2D contents in figure 4, the orientation of line AB changes. When we do this we bring

our perceptions into line with our beliefs – exactly what Fodor and Pylyshyn seek to deny by

their use of lower-level illusions. Thus, to the extent that bottom-up picture perceptions are

visually revisable at personal-level of awareness – with the exception of trompe l’oeils, which I

will argue constitute rule-proving exceptions (section 5) – such picture perceptions do not

constitute lower level illusions.

Perhaps certain depiction theorists can come to the rescue of cognitive impenetrability. Some

accounts of depiction hold that we have experiences of pictures as 2D and 2.5D simultaneously

(Wollheim 1987, 46; Gibson 1986, 282; Rock 2001, 98), which contradicts the account given

above of figure 4. In its extreme form – the claim that we always visually experience pictures as

24

2D and 2.5D simultaneously – this proposal (which consists essentially of phenomenological

claims) is in fact contradicted by the phenomenology: in bottom-up picture perceptions, which

are the most frequent kind, the too experiences are separated by the considerable mental effort

required to switch from one of the two visual experiences to the other. But in fact, even to say

that we sometimes have a visual experience both of the surface and the picture-content seems

wrong. In order to see a cube drawing as both flat and cubical at the same time, we would have

to represent the intersection of the enclosed lines (point A) as occupying two locations at once.

There are several ways of generating visual contents from pictures, but neither the natural, hard-

wired ways nor the attention-driven visual interpretations make us represent a point at two

different locations simultaneously. Although the visual system can yield separate and conflicting

depth interpretations of scenes, each interpretation always seems to be consistent (Waltz 1975;

Pylyshyn 1993, 99-107; Cutting & Massironi 1998 also give several illustrations of this visual

assumption in their discussion of line interpretation). When vision cannot solve the problem of

depth relations in a scene consistently, the result is an incomplete content and visual experience

of only part of the picture at a time, as when we see the devil’s pitchfork or certain drawings by

Escher. Moreover, it is not part of the phenomenology of the picture-perception that we see the

point in two places at once; on the contrary, it is very much part of the phenomenology that we

see the point at one location or the other and that the two experiences are separated temporally

by an ‘attentional switch’ (which itself seems to be positively experienced). In fact, in the case of

the simple outline drawing of a cube, there is little else to the phenomenology of the perception

than the locating of these points in space.

25

In that case, what could explain the insistence of Wollheim, in particular, on the twofoldness

thesis? Wollheim holds that the twofold experience of pictures is a single experience with two

‘aspects’. Perhaps we could give the following construal of that claim. Matthen (2005, 309-313)

has argued that pictures do not engage the motion-guiding visual system, a visual system thought

to be implemented in the dorsal pathway and to be distinct from the system dedicated to

recognition and to representing objet features (Ungerleider & Mishkin 1982; Milner & Goodale

1995). According to Matthen, the impact of this physical difference on the phenomenal character

of picture perceptions is that they do not cause a ‘feeling of presence’ (2005, 306). So perhaps

the dorsal system dedicated to navigation ‘knows’ that the picture is a more or less flat object,

while at the same time the ventral system picks up the volumetric contents and depth relations

from the picture’s surface. This would solve the apparently paradoxical nature of

phenomenological accounts of twofoldness found not only in Wollheim (1987, 46) but also in

Peacocke (1987, 386, 394) and in Gibson (1986, 282). Wollheim insists that the viewer ‘remains

visually aware not only of what is represented but also of the surface qualities of the

representation’ (1980, 216, italics added), even describing the experience as ‘twofold attention’.

Perhaps this can be squared with the kind of twofoldness patched together in the last paragraph.

Even so, a problem would remain: there is no room in this account of twofoldness for the fact

that we consciously perform mental actions in order to switch from one set of visual contents to

the other. If we really are visually aware of two contents at once, as Wollheim holds, they cannot

be the two sets of contents that we have trouble switching to and from when we perceive figure 4

(or, for that matter, figures 2 and 3).

26

Another attempt to explain Wollheim’s convictions about twofoldness could be that they result

from conflating two senses of ‘seeing’, a contentful and a purely causal sense: the surface causes

the picture content and in this sense we see it, but when it does this, we don’t see the surface

where it is – since visually experiencing the content means experiencing different points of the

surface at different distances. Once again, this does not account for Wollheim’s particular

position, since he insists that we do not see the surface in the bare sense that the surface is a

causal stimulus, but that we visually experience it. Yet another explanation could be this, which

concerns colour in particular. The colour that we place as a feature at a location, usually behind

the picture-surface, is the colour caused by the surface. So one may hold that we have an

accurate perception of the surface in respect of colour and at the same moment an inaccurate

perception in respect of location. The location-component of the content is inaccurate because

there is no blue there, while the feature-component is accurate because there is blue (but not

where it is located). In that case, we do not see the colour in two places at once: we see it behind

the picture-surface. This account would be compatible with typical phenomenological

descriptions of perceiving pictures in Wollheim, as when he says that we ‘marvel endlessly at the

way in which line or brush stroke or expanse of colour is exploited to render effects or establish

analogies that can only be identified representationally’ (Wollheim 1980, 216). However, the

solution cannot be extended to shapes. Colours (including non-chromatic ones in etchings and

drawings) yield shape information, but adopting the solution for shapes would again lead to the

problem encountered earlier, of seeing the same point in two places at once. Thus, it’s very

tempting to conclude that the position defended by Gombrich and Cutting is right: we cannot

perceive both the surface and the content of pictures at the same time.

27

The upshot of this discussion of simultaneity and twofoldness (theses to the effect that we have

visual experiences of pictures as 2D and 2.5D simultaneously) is that they do not offer a

plausible alternative to the account given above of visual ambiguity in bottom-up, naturalistic

picture perceptions. Such pictures can admit a form of visual ambiguity which implies that

certain visual processes which lead from the 2D sketch to the 2.5D sketch are penetrable by

consciously driven processes. This modulation of visual processes does not just change which

data is processed (as occurs in cases of attention-shifting) but the way visual data is processed. I

think it unlikely that all bottom-up picture perceptions are subject to this form of ambiguity, but

it suffices that many of them are to establish the penetrability of the relevant visual processes.

The robustness of the picture contents caused by naturalistic pictures – the difficulty we have in

generating visual experiences with 2D contents while looking at naturalistic pictures – favours

depiction and is an enabling condition for pictorial representation. For picture-perception as we

know it to occur, we have to be able to continue to exploit the depth representations formed by

earlier cognitive stages even after we have reached the stage at which we can form doxastic

states. Otherwise, the mental representation of what the picture represents would evaporate each

time we stopped making an effort to direct spatial attention in the right way. Imagine the

opposite scenario. If the 2.5D visual interpretation did not prevail naturally each time we looked

at the picture, perceiving pictures would be a painfully slow and attention-consuming activity;

we would only be able to understand one part of a picture at a time with considerable effort. In

that case, seeing the opaque cube-drawing as a cube would be as hard to do as it is now to see the

cube-drawing as a flat surface. Picture perception would require great mental concentration on

one region of the picture at a time. Higher-level attention is a valuable and energy-consuming

28

resource for the brain, and thus it is unlikely that depiction would be as widespread an activity as

it actually is. Instead, the 2.5D visual interpretations prevail, and depiction as we know it –

especially the profusion of different individual styles in drawing and painting – has emerged by

exploiting that robustness. At the same time, admitting cognitively driven picture perceptions

alongside bottom-up ones (see the next section) can explain both why we can use found objects

as fortuitous pictures, and how we understand picture styles that exploit or frustrate visual

procedures in different ways, either for aesthetic effect or due to technical limitations.

4. Cognitively driven picture perceptions

In the second form of visual ambiguity supported by pictures, the picture surface causes a

veridical visual experience with 2D content in a bottom-up way by the time we reach

phenomenal awareness; but it can subsequently support a visual experience of a 2.5D scene if we

agentively control spatial attention, optionally with backup from conceptually encoded

information. The less ambiguously and more fluently a surface causes a 2D visual representation,

the greater the conscious effort required to use the surface to have an experience with volumetric

and depth content. For example, in figure 5a, the visual system detects no occlusion. It is

unlikely to detect a surface discontinuity because the change in colour is not sharp or regular

enough. In the absence of other depth cues, the object is likely to suggest to the visual system a

roughly plane, irregularly stained surface. (With some priming, however, we can see the figure

as a concavity lit from the left.) Figure 5b supports an experience of the lower and upper halves

as foreground and background (see section 1). If we have trouble performing this segregation, it

helps to think of the figure as a field with stalks against a gray sky. If the same object is viewed

29

upside-down (5c), the figure-ground effect goes away, possibly because of an implicit

assumption that light comes above (see Ramachandran 2004). But in that case, we can reinstate

the figure-ground segregation and the nonconceptual depth representation if we are given

conceptual cues; eg, if we’re told to see the figure as a sandy stretch in the foreground with reeds

or shrubs in the background.

a b c

Figure 5

Both the phenomenal character and the nonconceptual content of visual perception change when

the figures are seen as either segregated figure and ground or as a convexity, instead of as a

stained flat surface. According to theories that defend perceptual impenetrability (Pylyshyn

1999; Pylyshyn 2007, 72; Raftopoulos 2009, 134), the processes that construct the

nonconceptual volumetric representations in Marr’s 2.5D sketch are part of early vision and thus

claimed to be impenetrable. Perceptual representations of depth are part of Marr’s 2.5D sketch.

30

(The 2.5D sketch is no less three-dimensional than the 3D sketch; the difference is that the 2.5D

sketch is ‘a viewer-centered representation of the depth and orientation of the visible surfaces’,

the 3D sketch a representation whose ‘coordinate system is object centered’; Marr 2010, 330.)

But because conceptual information can determine which nonconceptual depth representations

we have of the figures, this activity of visualizing or seeing-in is not just a case of perceptual

ambiguity which can be contained within lower-level processes. Instead, it fits Pylyshyn’s

definition of cognitive penetrability: the visual experience of depth is ‘altered in a way that bears

some logical relation to what the person knows’ (Pylyshyn 1999, 343) – namel, to conceptual

information or memories of the textures of specific kinds of scenes. Thus, the processes that

yield volumetric representations as outputs turn out to be cognitively penetrable.

Note that for figure 5b, while we can have conceptual representations (a field of wheat against a

gray sky), perhaps we can also have a nonconceptual depth representation without having to use

any concepts, solely on the basis of the figure-ground segregation. However, in that case, the

segregation would not have to be performed bottom-up; it is possible to perform it consciously

by imagining that there is a distance depth-wise between the top and bottom parts of the figure.8

This would count as a cognitively driven (personal-level, though not conceptual) form of amodal

completion. In Kanizsa and Gerbino’s terminology, the depth perception in figure 5c – and

perhaps 5b, depending on the subject’s perceptual set – could be described as a case of

8 If we perform the figure-ground segregation, do we have the same visual experience of figure

4b with and without the concept? Siegel 2006, 2011a, thinks we do not; Lyons (2011, 305)

thinks that the conceptual phenomenology may be ‘a late experiential effect, leaving the

nonconceptual early perceptual states unaffected but influencing the nondoxastic seemings’.

31

represented completion as opposed to perceptual completion. As Kanizsa and Gerbino put it,

such a construal of the figure does not give us

the impression of being faced with something “objective,” independent of us, not

influenced by our will or our cognitive set. Indeed, properties such as phenomenal

givenness and independence from the observer characterize a perceptual datum and

distinguish it from a datum that is merely thought (Kanizsa and Gerbino 1982, 174).

Briscoe (2011) also calls similar completions ‘cognitive’; and elsewhere uses the term ‘make

perceive’ to describe comparable top-down cognitive activities: ‘one engages in make-perceive

when one projects or “superimposes” a mental image on a certain region of the visually

perceived world’ (Briscoe 2008, 482).

Like the claim about penetrability made in the last section, this one too is of a different kind to

the claims made in Macpherson 2006. Macpherson discusses cases of perceptual ambiguity in

which she argues that the phenomenal character of perception changes but the nonconceptual

content can remain the same, whereas the cases of seeing-in described here bring about changes

in content along with phenomenal character. This is because the cases discussed in Macpherson

2006 essentially concern mental rotations of represented shapes without altering those shapes,

whereas the cases described here concern alteration of the shapes themselves (from 2D to 3D or

vice versa). On the other hand, the form of penetrability described may be compatible with part –

32

but not all – of Macpherson’s 2011 account of the penetrability of visual processes by the

imagination.9

The need for top-down cognitive influence in picture perception is not limited to fortuitous

pictures. Paintings and drawings in many styles require subjects to make an active personal-level

effort to visually experience their contents. This can be as a result of technical limitations like

absent or imperfect perspective (in Roman wall paintings for instance), or deliberate choice in

caricature and in partly abstract styles (like expressionism, cubism, and so on). Many of Turner’s

paintings, or parts of the paintings (of which the above figures are photographs) are in this

category. While it is possible to perceive depth in those paintings, Ruskin held that Turner’s

paintings teach the eye to be ‘innocent’ about content and focus on traits of the picture surface;

for Ruskin, as Gombrich put it, ‘we do not even see the third dimension, only patches of colour

and textures’ (1960, 238).

5. Picture perceptions and perceptual beliefs

To claim that cognitive and agentive states can modulate the outputs of visual processes is

potentially to admit circularity into the justificatory relation between perceptual experience and

belief. As Siegel (2011b) puts it, if Jill visually experiences Jack’s face as angry because she

9 That is, I do not agree with Macpherson that Delk & Fillenbaum’s and Hansen & Olkkonen’s

experiments show colour perception to be cognitively penetrable; see Zeimbekis 2012b. But part

of her account of the imagination, when applied to shapes as opposed to colours, may involve

cognitively driven visualizings.

33

believes that Jack is angry, then visually experiencing Jack’s face as angry can no longer justify

her belief that Jack is angry. Siegel’s question applies in the following way to the forms of visual

ambiguity outlined. On one hand, the experiences generated when we see figure 4 as flat, and

when we see figures 5a-c as 2.5D, are in some sense visual experiences. At the same time, those

experiences are caused jointly by us – by agentively driven activities jointly with bottom-up

processes. The suggestion is that we can contribute to deciding which visual experiences we

have, and this would vitiate the ability of visual experience to justify belief.

Note that, in one sense, it is relatively trivial that we can decide which visual experiences to

have. If we accept that we can never experience simultaneously all of the nonconceptual visual

contents that a visual scene could cause, because we cannot focus attention on all parts of a

visual scene at once,10

then, given that we are capable of agentively directing the focus of visual

attention, it follows trivially that we choose what to visually experience. For example, if

applying the concepts pine and tree to the same object required having different visual

experiences of it, that would not require having contradictory contents and would not vitiate the

justificatory relation: the visual experience which supports the concept pine would also support

the concept tree; the concepts name kinds standing in a determination relation. But that is not

what happens in cases of visual ambiguity – at least not in the ones described here.11

The

ambiguity of figures 4 and 5 allows us to have contradictory visual contents, so the threat is that

those visual contents can support contradictory attributions or beliefs.

10

For a description and plausible defense of this position see Dehaene 2007.

11 It does happen in Siegel’s example of perceptions of pine trees.

34

There are two ways to block the inference from cognitive penetrability to the vitiation of

perceptual justification. The forms of penetrability described afflict picture perception, so one

response would be to claim that even if processes which generate volume perception in picture

perception are cognitively penetrable, volume perception processes in object perception are not

penetrable. Another response would be to show that of the two states caused in each case of

visual ambiguity, at most one state has the roles of causing and justifying beliefs – in other

words, that only one state counts as a perception functionally and epistemically. In fact, these

two questions turn out to be very closely related. I’ll start from the second question, and this will

give us the means to settle the first question.

If the capacity of visual perception to justify beliefs is vitiated by either of the forms of visual

ambiguity outlined, it is unlikely to be the second form, in which visual processes naturally yield

a veridical perception of the picture’s surface as flat and we subsequently use the surface as a

visual prop to visualize a 3D scene. It immediately seems suspect to suggest that a cognitively

driven act of visualizing could compete as a state for perception’s causal and justificatory roles.

When a fortuitous picture, like the wall stains described by Leonardo, naturally causes a visual

experience of a flat surface, that experience causes a perceptual belief with the same content.

Visual experiences of depth are produced after the formation of that perceptual belief by

agentively controlling the focus of spatial attention. They are something we do akin to a mental

action or act of visualizing or the imagination, not something that happens to us, that is, the

outcome of purely stimulus-driven and hard-wired processes. Those states do not feel like

perceptions and there is no reason to think that they have the causal role of causing perceptual

beliefs which could compete with the belief caused by the initial, veridical visual experience.

35

Matters are more complicated when it comes to penetration of visual processes which support

naturalistic picture perceptions. Remember that in naturalistic picture perceptions – much as in

object perception – by the time we reach phenomenal awareness, we have already formed a

mental representation of a 2.5D scene. But the processes which cause the 2.5D representation

turn out to be cognitively penetrable: with mental effort, we can also visually experience the

picture or parts of it as flat. I will try to show that the 2.5D visual experience does not cause

beliefs and does not count as a perception. In other words, the fact that such pictures support

cognitive penetrability does not mean that they support epistemic situations in which we can

justify conflicting beliefs by using different sets of visual contents.

The reason why picture perceptions do not cause higher-level illusions – beliefs whose contents

are the contents of the picture perception – is that they are not the same kinds of states as object

perceptions. While the states and experiences of bottom-up picture perceptions ‘free-ride’ on

object perception processes and have much on common with them, that does not mean that they

are identical to them. The key difference is that in picture perception the brain does not calculate

volumes and depth relations by using either binocular disparity from stereopsis or parallax from

movement. According to Marr (2010, Ch. 3), stereopsis is largely responsible for generating the

‘viewer-centered representation of the depth and orientation of the visible surfaces’ (330) in

object perception – the 2.5D sketch which, as we saw, is likely to provide the nonconceptual

content of visual perception. Picture perception, on the other hand, can exploit neither stereopsis

nor parallax, and relies exclusively on the kinds of monocular depth cues described in section 2.

Therefore, binocular picture perception generates 2.5D brain representations, and the

36

nonconceptual contents of phenomenal awareness, not just by using monoptic perception, but by

using a subset of the processes of monoptic object perception: not only is stereopsis absent, as it

is in monoptic vision, but parallax from movement, which is available to monoptic object

perception, is also absent from picture perception. As brain states, picture perceptions are very

different states to object perceptions.

These differences between the two kinds of visual states seem to be reflected by differences in

their phenomenology prior to the onset of belief – that is, by differences in their phenomenal

character. Briscoe (2008) has pointed out that the case of Susan Barry (reported in Sacks 2010),

who had monoptic vision and only experienced seeing stereoptically for the first time at an

advanced age, suggests that the visual experiences caused by stereoptic and monoptic vision can

be distinguished phenomenally:

I noticed the edge of the open door to my office seemed to stick out toward me.

Now, I always knew that the door was sticking out toward me when it was open

because of the shape of the door, perspective and other monocular cues, but I had

never seen it in depth. It made me do a double take and look at it with one eye and

then the other in order to convince myself that it looked different. It was definitely

out there. […] While I was running this morning with the dog, I noticed that the

bushes looked different. Every leaf seemed to stand out in its own little 3-D space.

The leaves didn’t just overlap with each other as I used to see them. I could see the

SPACE between the leaves. The same is true for twigs on trees, pebbles on the

37

road, stones in a stone wall. Everything has more texture. (Sacks 2010, 125; quoted

by Briscoe 2008, 477)

Susan Barry’s reports of the phenomenology of her visual experiences suggest that depth is

represented much less vividly in monoptic object perceptions than in stereoptic ones – and this,

despite the fact that parallax was available to the subject, unlike in picture perceptions. She also

reported that before her vision was corrected, she predicted that she could imagine what

stereoptic vision was like, but had to retract the claim after she experienced stereopsis (Sacks

2010, 122), which suggests that stereopsis acquainted her with a new kind of experience. Briscoe

describes the difference between seeing monoptically and seeing stereoptically as a ‘dramatic

influence of binocular depth information on the spatial phenomenal character of our visual

experience’ (2008, 477). His explanation of the phenomenal difference is that after the brain

calculates three-dimensional shape on the basis of binocular disparities in the light information

reaching the retina, the information and processes are ‘not lost in our conscious visual experience

of the object. Indeed […] we can literally see the difference made by their presence (and

absence) in the light available to the eyes’ (Briscoe 2008, 477).

The difference between object perceptions and picture perceptions is that between (a)

constructing the depth representations available to phenomenal awareness by means of

stereopsis, parallax, and hard-wired processes that interpret monocular cues, and (b) constructing

those representations only by means of monocular cues. Susan-Barry-type experiences are

subserved by an intermediate state, (c): parallax plus monocular cues, but no stereopsis. Since

there is a phenomenal difference between (a) and (c) that a subject whose physiology is working

38

in the relevant respect (like Susan Barry after her vision was corrected) is able to distinguish,

there is all the more reason why a subject should be able to tell apart (a) normal object perception

from (b) picture perception.

The hypothesis about stereopsis and parallax can be supplemented with other hypotheses about

the difference between object and picture perception. As already mentioned in section 3, Matthen

(2005, 309-313) has argued that pictures do not engage the motion-guiding visual system, but

only the pathway dedicated to recognition and to representing objet features (the distinction is

made by Ungerleider & Mishkin 1982, and Milner & Goodale 1995). Matthen claims that the

impact of this physical difference on the phenomenal character of picture perceptions is that they

do not cause a ‘feeling of presence’ (2005, 306). Another feature of picture perception which is

compatible with both the lack of stereopsis and parallax, and the lack of engagement of the

motion-guiding system, could be the lack of subject independence. Subject independence is

claimed by Siegel (2006b) to be part of the phenomenal character of object perceptions. This

feature is absent from picture perceptions not only because the locations of objects represented in

picture perception are insensitive to parallax, but also because, as Cutting as shown, shapes

represented in picture contents withstand considerable changes in the viewing angle (Cutting

1987).

These differences between object and picture perception are at the level of the states which bear

the content, and not – or at least not only – at the level of the contents of the states.12

The lack of

12

Lack of stereopsis does seem to have an impact on representational content itself; for example,

it leads to miscalculations about the shape and location of objects (Sacks 2010, Ch. 5). However,

39

subject-independence cannot be detected as a change in content, since subject-dependence means

precisely that egocentric content does not change as it should along with viewing position. Lack

of engagement of the motion-guiding system has functional effects, not effects on the object

features represented visually. Something similar applies to the suppression of stereoptic

representations of depth. When reporting her experiences, Susan Barry contrasts representing the

orientation of a door ‘because of the shape [and] perspective’ to ‘seeing it in depth’, and

representing the ‘overlap’ of leaves to ‘seeing the space between the leaves’. But the locations

behind an occluding leaf are not seen in stereoptic vision either; and the depth-wise angle of the

door is calculated and represented on the basis of phenomenal shape and perspective in monoptic

vision (just as it is in pictures). So stereoptic seeing does not necessarily add something to the

content of a visual state; the difference may be in the mode or way that the content is

represented, the quality of the state itself, or epistemic feelings. Finally, the absence of the brain

states and processes that calculate disparity and parallax, the lack of engagement of the dorsal

visual system, and increased reliance on monocular cues, may not affect only the

phenomenology of picture and object perceptions: it may also mean that the causal antecedents

for perception to play its functional role of causing perceptual beliefs are not satisfied, something

which could directly prevent picture perceptions from causing perceptual beliefs towards the

picture contents.

the difference at the level of content may not be detectable epistemically, that is, by the subjects

having the contents; it seems to be a condition for having misrepresentational content that it

should not be detected as being misrepresentational. The differences described by Susan Barry

were not noticed by comparing monoptic and stereoptic perceptions of the same scene; they

seem to have been detected as overall changes in the quality of her experience.

40

This account of the causal and epistemic role of picture perceptions is supported by a

consideration of pictures that do cause illusions: trompe l’oeils. ‘Trompe l’oeil’ is an ambiguous

term which sometimes designates a picture and sometimes the effect of a picture. Here, I use the

term for pictures in contexts in which they succeed in causing higher-level illusions, typically

signaled by a sense of surprise when the illusion is dispelled. (I do not use the term for paintings

that are extremely naturalistic but appear in contexts or conditions in which they do not produce

the illusion, like William Harnett’s Old Models; see Goldstein 2005, 354.) Figure 6, while it is

only a photograph of a picture (a mural) by John Pugh, provides a good illustration.

Figure 6

Trompe-l’oeils, for as long as they work, make us ascribe a property represented by the picture to

a location in the space represented by the picture. The picture contents – nonconceptually

represented shapes, colours and textures, concepts for properties and kinds, and object files for

objects to bind the features and instantiate the properties – reach doxastic-level awareness

41

wholesale, and form the structured contents of a perceptual belief. We could form such a belief if

we turned a corner and suddenly saw the building in figure 6. That in turn implies that the

phenomenal character of the experience, and the causal role of the state, are indiscernible from

those of an object perception.

Trompe-l’oeils do not get that far under any contextual conditions. A precondition for their

success is that the picture should not be presented as a picture, something which would affect

‘perceptual set’: the preparedness for certain categories of object, scene or action, which is

capable of biasing the visual processing of depth relations such as figure-ground segregation and

could therefore defeat the trompe l’oeil’s ability to cause an illusory belief (see Vecera 2000,

367-370, for an overview of the concept of perceptual set, and Peterson and Hochberg, 1983, for

the claim that perceptual set can affect figure-ground segregation). This is why successful trompe

l’oeils usually appear in particular visual contexts, which could be described as unexpected in

one sense, but also – in another sense – as consistent with expectations. On one hand, they

benefit from being embedded in contexts where we do not expect to see a picture, which is why

they are often painted on architectural features like external or internal walls, ceilings and

columns. On the other hand, the content of the trompe l’oeil should be consistent with the

context in which it is embedded, which is why contents frequently consist of fluting, panels,

mouldings, and other features consistent with the architectural setting. When such contextual

conditions are met, even coarsely made pictures which are not naturalistic can momentarily fool

the mind into confusing the picture perception with an object perception.

42

But the condition which is of direct interest here is that trompe l’oeils work as long as the

absence of binocular disparity can go undetected and as long as parallax is neutralized, if we are

not stationary. Thus, the ability of the picture to cause the illusionistic effect increases with the

viewer’s distance from it, and the distance required is itself proportionate to the depth

represented in the picture contents: the shallower the scene represented, the smaller the binocular

disparities it would cause and the harder it is to detect their absence. A frequently used theme in

trompe l’oeils are false window casings painted high on a building, which meet all of these

conditions, as well as the earlier set of conditions connected with perceptual set. When trompe

l’oeil illusions are dispelled, it is usually because the absence of parallax is detected as we move

relative to the wall, column or ceiling in which the picture is embedded.

Once the illusion is dispelled, the experience of the picture seems to change. For example, when

we realize that the lines suggesting the presence of a window casing are planar and that there are

no window casings where we represent them, we can still mentally represent the casings – in

other words, the content of the picture – as we look at the building. Support for this claim comes

from the phenomenon that Kubovy (1986) calls ‘robustness of perspective’: picture perception is

relatively insensitive to changes in the angle from which the picture is viewed; according to

Cutting (1987), pictures can be rotated up to 22 degrees (on their vertical axis) without this

deforming the three-dimensional representations they cause. This is a very substantial degree of

rotation, so we should be able to represent picture contents throughout the movements that make

us detect the lack of parallax. What changes throughout the movements is the illusory nature of

the picture contents, not the contents themselves. The fact that picture perceptions and trompe

43

l’oeil illusions represent the same contents under different psychological modes or attitudes

corroborates the claim that picture perceptions are distinct from perceptual illusions.

Therefore, trompe l’oeils are rule-proving exceptions to the theory that picture-perceptions and

object-perceptions have different phenomenal characters. They are exceptions because they

produce picture perceptions which are momentarily indiscernible from object perceptions. But

they are rule-proving exceptions because what makes them indiscernible from object perceptions

is the fact that we do not detect the absence of binocular disparity and parallax. As such, they

confirm the thesis that lack of stereopsis and parallax underlie the phenomenal difference

between the remaining cases of bottom-up picture perception, and object perceptions.

6. Conclusion

Let me summarize the last section’s argument. We can decide which visual experiences to have

of pictures by consciously affecting how the processes that generate a fundamental part of our

visual contents – volumetric shape and depth relations – are carried out. This suggests that we

can contribute to deciding which visual experiences we have; something which would vitiate the

ability of visual experience to justify belief. Here, I have accepted the conclusion about

penetrability as plausible, but blocked the inference from penetrability to vitiation of the

perceptual justification of belief. I have tried to show that the states resulting from cognitive

penetration of visual processes have neither the right functional roles for them to cause beliefs,

nor the right epistemic roles for subjects to use them to justify beliefs. In effect, I have argued

44

that neither form of what we have been calling picture perception – neither the hard-wired nor

the cognitively driven kind – is really a form of perception in the functional or causal sense.

Could we also conclude that the processes which generate volume perception are cognitively

penetrable only during picture perception, not in the wider context of the processes that take

place in object perceptions?

Think of object perceptions on a range. At one end are cognitively driven picture perceptions;

next are pictures with weak depth cues; after that, naturalistic pictures with good depth cues,

followed by trompe l’oeils; and finally, object perceptions in good viewing conditions. Object

perceptions would be even harder to visually experience as 2D than trompe l’oeils, which are

already very hard to visually experience as 2D. Stereopsis, and visual exploration using parallax,

would make it so difficult to have such visual experiences that we would have to actively

visualize the scene before us as being 2D. If it was possible to have such a state, it would have to

be by keeping still to prevent the effect of parallax and viewing the object through one eye to

stop the brain calculating binocular disparity; and even then, we would still have monocular cues

to contend with, so such a state could still not emerge solely on the basis of stimulus-driven and

hard-wired processes.

Now, if we could not get to the point of having such a state, the processes which generate

volume perception would not be cognitively penetrable in visual object perception taken as a

wider set of visual processes than picture perception; they would only be penetrable in picture

perception. This would be an interesting outcome, because challenges to impenetrability and

45

modularity from visual ambiguity always use pictures as examples: pictures are used to show

that there is penetration of some process, and then the conclusion is that that process is generally

penetrable. The conclusion would be false if in object perception we simply could not get

ourselves into visual states whose contents were two-dimensional.

But suppose that we could have such states. It is plausible that we can, because viewing

conditions can be far from optimal in object perception (especially for distant objects, for the

reasons seen in the discussion of trompe l’oeils), and the visual system is made to be able to deal

with such conditions (just as it seems to do when we can make sense of unnaturalistic pictures).

Then – for the same reason that cognitively driven picture perceptions do not cause perceptual

beliefs – these consciously sustained mental representations of real scenes as flat would not

cause perceptual beliefs either. They would be visual experiences that we ourselves actively

sustain as agents by performing acts of visualizing, not states that happen to us like purely

stimulus-driven and hard-wired visual contents are. Yet, it would remain that the processes

which generate volume perception in object perception are cognitively penetrable; it is just that

the states resulting from penetration would not cause beliefs, and could therefore not vitiate the

justification of beliefs by visual perception.

In either case – whether the visual processes that construct volume and depth are penetrable only

in picture perception, or also in object perception – it transpires that we can admit the cognitive

penetrability of fundamental visual processes without threatening the epistemic relation between

visual perception and belief. So we do not have to deny cognitive penetrability to uphold the

perceptual justification of belief.

46

References

Abell, C. & Currie, G. 1999. Internal and External Pictures. Philosophical Psychology 12.4: 440-

441.

Biederman, I. 1987. Recognition by components: a theory of human image understanding.

Psychological Review, 94, 115–147.

Biederman, I. 1995. Visual object recognition. Kosslyn, S. & D. N. Oshershon (Eds.), An

invitation to cognitive science. MIT Press. 121-165.

Biederman, I. & Kim, J. 2008. 17000 years of depicting the junction of two smooth shapes.

Perception 37: 161-164.

Block, N. 1990. Consciousness and accessibility. Behavioral and Brain Sciences 13, 596-598.

Block, N. 1997. On a confusion about a function of consciousness. In Block, N., Flanagan, O.

and Güzeldere, G. (Eds.) The Nature of Consciousness: Philosophical Debates, MIT Press.

Block, N. 2005. Two neural correlates of consciousness. Trends in Cognitive Sciences 9. 2, 46-

53.

Briscoe, R. 2008. Vision, action and make-perceive. Mind & Language 23: 457–497.

Briscoe , R. 2011. Mental imagery and the varieties of amodal perception. Pacific Philosophical

Quarterly 92.2, 153-173.

47

Bulthoff, H. & Edelman, S. 1992. Psychophysical support for a two-dimensional view

interpolation theory of object recognition. Proceedings of the National Academy of Sciences of

the United States of America 89: 60-64.

Cavanagh, P. 2005. The artist as neuroscientist. Nature 434.17: 301-307.

Churchland, P. M. 1988. Perceptual plasticity and theoretical neutrality: A reply to Jerry Fodor.

Philosophy of Science 55: 167–187.

Cutting, J. 1987. Rigidity in cinema seen from the front row, side aisle. Journal of Experimental

Psychology 13.3, 323-334.

Cutting, J. and Massironi, M. 1998. ‘Pictures and their special status in cognitive inquiry’. In

Hochberg, Perception and Cognition at Century’s End. Academic Press.

Delk, J. and S. Fillenbaum. 1965. Differences in perceived color as a function of characteristic

color. American Journal of Psychology 78.2, 290-293.

Edelman, S., Bulthoff, H.H., 1992. Orientation dependence in the recognition of familiar and

novel views of three-dimensional objects. Vision Research 32.12, pp. 2385-2400.

Fodor, J. 1975. The Language of Thought. Harvard University Press.

Fodor, J. 1983. The Modularity of Mind. MIT Press.

Fodor, J. 1988. ‘A Reply to Churchland's “Perceptual Plasticity and Theoretical Neutrality”’.

Philosophy of Science 55.2: 188-198.

Fodor, J. 1990. A Theory of Content and Other Essays. MIT Press.

Gibson, J. 1986. The Ecological Approach to Visual Perception. New York, Taylor and Francis.

48

Goldstein, B. 2005. ‘Pictorial perception and art.’ In Bruce Goldstein (Ed.), Blackwell Handbook

of Sensation and Perception, 344-378.

Gombrich, E. 1984 [1960]. Art and Illusion: a Study in the Psychology of Pictorial

Representation. London: Phaidon Press.

Hansen, T., Olkkonen, M., Walter, S. and Gegenfurtner, K. 2006. Memory modulates color

appearance. Nature Neuroscience 9.11, 1367-1368.

Kanizsa, G., & Gerbino, W. 1982. ‘Amodal completion: Seeing or thinking?’ In B. Beck (ed.),

Organization and Representation in Perception, 167-190. Hillsdale, N.J.: Erlbaum.

Kosslyn S. 1994. Image and Brain: The Resolution of the Imagery Debate. Cambridge, MA:

MIT Press.

Kourtzi, Z., Grodd, W., and Bulthoff, H. 2003. Representation of the Perceived 3-D Object

Shape in the Human Lateral Occipital Complex. Cerebral Cortex 13: 911-920.

Kubovy, M. 1986. The psychology of perspective and Renaissance art. London, Cambridge

University Press.

Lamme, V. 2000. Neural Mechanisms of Visual Awareness: A Linking Proposition. Brain and

Mind 1: 385-406.

Lamme, V. 2003. Why visual attention and awareness are different. Trends in Cognitive Sciences

7.1, 12-18.

Lamme, V. 2005. Independent neural definitions of visual awareness and attention. In A.

Raftopoulos (Ed.), The Cognitive Penetrability of Perception: An Interdisciplinary Approach.

Nova Science, New York.

49

Leonardo da Vinci. 1989. On Painting. Yale University Press.

Levinson, J. 1998. Wollheim on Pictorial Representation. Journal of Aesthetics and Art Criticism

56.3: 227-233

Logothetis, N.K., Pauls, J., Poggio, T., 1995. Shape representation in the inferior temporal cortex

of monkeys. Current Biology 5 (5), 552–563.

Lopes, D. 1996. Understanding Pictures. Oxford University Press.

Lyons, J. 2011. Circularity, reliability, and the cognitive penetrability of perception.

Philosophical Issues 21: 289-311.

Macpherson, F. 2011. ‘Cognitive penetration of colour experience: Rethinking the issue in light

of an indirect mechanism.’ Philosophy and Phenomenological Research (forthcoming). doi:

10.1111/j.1933-1592.2010.00481.x

Macpherson, F. 2006. ‘Ambiguous Figures and the Content of Experience’, Noûs 40.1: 82-117.

Marr, D. [1982] 2010. Vision. A Computational Investigation into the Human Representation

and Processing of Visual Information. MIT Press.

Matthen, M. 2005. Seeing, Doing and Knowing. A Philosophical Theory of Sense Perception.

Clarendon Press, Oxford.

Milner, A.D. & Goodale, M. A. 1995. The Visual Brain in Action. Oxford University Press.

Naccachea, L. and Dehaene, Stanislas. 2007. Reportability and illusions of phenomenality in the

light of the global neuronal workspace model. Behavioral and Brain Sciences 30: 518-520.

Olkkonen, M., Hansen, T., and Gegenfurtner, K. 2008. Color appearance of familiar objects:

Effects of object shape, texture, and illumination changes. Journal of Vision 8.5, 1-16.

50

Peacocke, C. 1987. Depiction. Philosophical Review XCVI: 383–410.

Peacocke, C. 1992. A Study of Concepts. MIT Press.

Peterson, M. and Hochberg, J. 1983. Opposed-set measurement procedure: A quantitative

analysis of the role of local cues and intention in form perception. Journal of Experimental

Psychology: Human Perception and Performance 9: 183-193.

Pylyshyn, Z. 1999. Is vision continuous with cognition? The case for cognitive impenetrability of

visual perception. Behavioral and Brain Sciences 22: 341–423.

Pylyshyn, Z. 2003. Seeing and Visualizing: It’s Not What You Think. MIT Press.

Pylyshyn, Z. 2007. Things and Places: How the Mind Connects with the World. MIT Press.

Raftopoulos, A. 2009. Cognition and Perception. How do Psychology and Neural Science

Inform Philosophy? MIT Press.

Raftopoulos, A. 2011. ‘Ambiguous figures and representationalism’. Synthese 181.3: 489-514.

Ramachandran, V. & Rogers-Ramachandran, D. 2004. ‘Seeing Is Believing.’ Scientific American

- Mind, January 2004: 100-101.

Sacks. O. 2010. The Mind’s Eye. New York: Knopf.

Schier, F. 1986. Deeper into Pictures: an Essay on Pictorial Representation. Cambridge

University Press.

Siegel, S. 2006a. ‘Which properties are represented in perception?’ In Gendler & Hawthorne,

eds., Perceptual Experience. Oxford: Oxford University Press.

Siegel, S. 2006b. Subject and object in the contents of visual experience. Philosophical Review

115.3: 355-388

51

Siegel, S. 2011a. The Contents of Visual Experience. New York: Oxford University Press.

Siegel, S. 2011b. ‘Cognitive penetrability and perceptual justification.’ Noûs, forthcoming.

Tarr, M.J., 1995. Rotating objects to recognize them: a case study of the role of viewpoint

dependency in the recognition of three-dimensional objects. Psychonomic Bulletin and Review

2.1: 55-82.

Tarr, M.J. & Bulthoff. 1998. Image-based object recognition in man, monkey and machine.

Cognition 67, 1-20.

Tarr, M.J. & Pinker, S., 1989. Mental rotation and orientation-dependence in shape recognition.

Cognitive Psychology 21.28: 233-282.

Tye, M. 1991. The Imagery Debate. Cambridge, MA: MIT Press.

Ullman, S., 1998. Three-dimensional object recognition based on the combination of views.

Cognition 67, 21-44.

Ungerleider, L. & Mishkin, M. 1982. Two Cortical Visual Systems. In Ingle, D. J., Goodale, M.

& Mansfield, R. J. W. (Eds.), Analysis of Visual Behaviour. Cambridge, Mass: MIT Press. 549-

586.

Vecera, S. 2000. ‘Toward a Biased Competition Account of Object-Based Segregation and

Attention’, Brain and Mind 1: 353-384.

Walton, K. 1984. Transparent pictures: On the nature of photographic realism. Critical Inquiry

11: 246-276.

Walton, K. 1990. Mimesis as make-believe: on the foundations of the representational arts.

Harvard University Press.

52

Walton, K. 1997. On pictures and photographs: objections answered, in Allen and Smith (Eds),

Film Theory and Philosophy: 60-75. Oxford.

Walton, K. 2002. Depiction, Perception, and Imagination: Responses to Richard Wollheim.

Journal of Aesthetics and Art Criticism 60.1: 27-35.

Waltz, D. 1975. Understanding line drawings of scenes with shadows. In P. H. Winston (ed.),

The Psychology of Computer Vision, 19-91. New York, McGraw-Hill.

Wollheim, R. 1980. Art and Its Objects. Cambridge University Press.

Wollheim, R. 1987. Painting as an Art. Princeton University Press.

Wollheim, R. 1998. On Pictorial Representation. Journal of Aesthetics and Art Criticism 56.3:

217-226.

Zeimbekis, J. 2012a. Pictures, Perception and Meaning. Manuscript.

Zeimbekis, J. 2012b. Color and cognitive penetrability. Philosophical Studies, forthcoming.